[WIP] Juno do not merge

common/config.yaml

@@ -4,8 +4,11 @@ jupyterhub::jupyterhub_config_hash:
   SlurmFormSpawner:
     start_timeout: 900
 
-profile::freeipa::mokey::require_verify_admin: false
+profile::freeipa::mokey::require_verify_admin: true
 profile::slurm::base::slurm_version: '23.02'
+# when using snapshots, it is quick enough to boot nodes that 900 seconds is enough for suspend
+profile::slurm::base::suspend_time: 900
+profile::gpu::install::passthrough::nvidia_driver_stream: '550-dkms'
 
 prometheus::global_config:
   scrape_interval: '1m'
@@ -22,3 +25,4 @@ prometheus::remote_write_configs:
     basic_auth:
       username: 'cqformation'
       password: "%{alias('prometheus_password')}"
+

mcgill-scs/config.yaml

@@ -0,0 +1,29 @@
+jupyterhub::jupyterhub_config_hash:
+  SbatchForm:
+    runtime:
+      min: 3.5
+      def: 3.5
+      max: 5.0
+    nprocs:
+      min: 1
+      def: 1
+      max: 1
+    memory:
+      min: 1024
+      max: 2048
+      def: 2048
+    oversubscribe:
+      def: true
+      lock: true
+    gpus:
+      def: 'gpu:0'
+      choices: ['gpu:0', 'gpu:1g.5gb:1', 'gpu:3g.20gb:1' ]
+    ui:
+      def: 'lab'
+  SlurmFormSpawner:
+    disable_form: false
+
+profile::software_stack::lmod_default_modules: ['StdEnv/2023', 'gcc/12.3', 'openmpi/4.1.5', 'python/3.10', 'ipython-kernel/3.10']
+profile::freeipa::mokey::require_verify_admin: true
+profile::slurm::base::slurm_version: '24.05'
+

mcgill-scs/custom.tf

@@ -0,0 +1,37 @@
+locals {
+  custom = {
+    nnodes = {
+      cpu = 0
+      cpupool = 0
+      gpupool16 = 6
+      gpupool16-cq = 4
+      gpupool12 = 2
+      gpupool12-j = 8
+      gpupool80 = 0
+    }
+    home_size = 100
+    project_size = 500
+    scratch_size = 400
+    nb_users = 1
+
+    user_quotas = {
+      home = "1g"
+      project = "2g"
+      scratch = "4g"
+    }
+
+    mig = {
+      gpupool16-cq = { "1g.5gb" = 7 }
+      gpupool16 = { "1g.5gb" = 7 }
+      gpupool12 = { "3g.20gb" = 2 }
+      gpupool12-j = { "1g.5gb" = 7 }
+    }
+
+    image_cpu = "snapshot-cpunode-2024-R810.5"
+    image_gpu = "snapshot-gpunode-2024-R810.5"
+
+    config_version = "dc6b37f4d2c077a37d88bf4862ba57a09eed7213"
+  }
+
+  name = "mcgill-scs"
+}

mcgill-scs/main.tf

@@ -0,0 +1 @@
+../common/main.tf

nvidia-workshop/config.yaml

@@ -0,0 +1,54 @@
+jupyterhub::jupyterhub_config_hash:
+  SbatchForm:
+    runtime:
+      min: 3.5
+      def: 3.5
+      max: 5.0
+    nprocs:
+      min: 1
+      def: 1
+      max: 1
+    memory:
+      min: 1024
+      max: 2048
+      def: 2048
+    oversubscribe:
+      def: true
+      lock: true
+    gpus:
+      def: 'gpu:0'
+      choices: ['gpu:0', 'gpu:1g.5gb:1', 'gpu:3g.20gb:1' ]
+    ui:
+      def: 'lab'
+  SlurmFormSpawner:
+    disable_form: false
+
+profile::software_stack::lmod_default_modules: ['StdEnv/2023', 'nvhpc/23.9', 'cuda/12.2', 'openmpi/4.1.5', 'python/3.11', 'ipython-kernel/3.11']
+profile::freeipa::mokey::require_verify_admin: true
+#profile::slurm::base::suspend_time: 86400
+
+profile::users::ldap::users:
+  dummy_cours1:
+    count: 1
+    groups: ['def-cours1']
+
+  dummy_cours2:
+    count: 1
+    groups: ['def-cours2']
+
+  dummy_cours3:
+    count: 1
+    groups: ['def-cours3']
+
+profile::slurm::accounting::accounts:
+  def-cours1:
+    Fairshare: 1
+    MaxJobs: 1
+  def-cours2:
+    Fairshare: 1
+    MaxJobs: 10
+  def-cours3:
+    Fairshare: 1
+    MaxJobs: 10
+
+

nvidia-workshop/custom.tf

@@ -0,0 +1,46 @@
+locals {
+  custom = {
+    nnodes = {
+      cpu = 0
+      cpupool = 0
+      gpu = 0
+      gpupool16 = 16
+      gpupool16-cq = 4
+      gpupool12 = 4
+      gpupool12-j = 20
+      gpupool80 = 0
+    }
+    cluster_purpose = "internal"
+    home_size = 50
+    project_size = 50
+    scratch_size = 50
+
+    user_quotas = {
+      home = "1g"
+      project = "1g"
+      scratch = "1g"
+    }
+
+    instances_type_map = {
+      juno = {
+        gpu = "gpu12-120-850gb-a100x1"
+      }
+    }
+    mig = {
+      gpu = { "1g.5gb" = 7 }
+      gpupool16 = { "1g.5gb" = 7 }
+      gpupool16-cq = { "1g.5gb" = 7 }
+      gpupool12 = { "3g.20gb" = 2 }
+      gpupool12-j = { "1g.5gb" = 7 }
+      gpupool80 = { "2g.20gb" = 3 }
+    }
+
+    image_cpu = "snapshot-cpunode-2024-R810.5"
+    #image_gpu = "Rocky-8.10"
+    image_gpu = "snapshot-gpunode-2024-R810.5"
+
+    config_version = "dc6b37f4d2c077a37d88bf4862ba57a09eed7213"
+  }
+
+  name = "nvidia-workshop"
+}

nvidia-workshop/main.tf

@@ -0,0 +1 @@
+../common/main.tf

test-mc-infra-cours/config.yaml

@@ -21,7 +21,6 @@ jupyterhub::jupyterhub_config_hash:
                 disable_form: false
 
 profile::freeipa::mokey::require_verify_admin: false
-profile::software_stack::lmod_default_modules: ['StdEnv/2023', 'gcc/12.3', 'openmpi/4.1.5', 'python/3.10', 'ipython-kernel/3.10']
 
 profile::users::ldap::users:
         dummy_cours1:

test-mc-infra-cours/custom.tf

@@ -1,15 +1,19 @@
 locals {
   custom = {
-    ncpu = 0
-    ncpupool = 1
-    ngpu = 0
-    ngpupool = 1
+    nnodes = {
+      cpu = 0
+      cpupool = 1
+      gpu = 0
+      gpupool = 1
+    }
 #    home_size = 100
 #    project_size = 100
 #    scratch_size = 50
-    image_cpu = "snapshot-cpunode-2024-R810.4"
-    image_gpu = "snapshot-gpunode-2024-R810.4"
+    image_cpu = "snapshot-cpunode-2024-R810.5"
+    image_gpu = "snapshot-gpunode-2024-R810.5"
 
+    config_git_url = "https://github.com/mboisson/puppet-magic_castle.git"
+    config_version = "1b45e1f"
     volumes = {
       nfs = {
         home = { size = 100, quota = "1g" }